Motion corrected fetal body magnetic resonance imaging provides reliable 3D lung volumes in normal and abnormal fetuses

Abstract Objectives To calculate 3D‐segmented total lung volume (TLV) in fetuses with thoracic anomalies using deformable slice‐to‐volume registration (DSVR) with comparison to 2D‐manual segmentation. To establish a normogram of TLV calculated by DSVR in healthy control fetuses. Methods A pilot study at a single regional fetal medicine referral centre included 16 magnetic resonance imaging (MRI) datasets of fetuses (22–32 weeks gestational age). Diagnosis was CDH (n = 6), CPAM (n = 2), and healthy controls (n = 8). Deformable slice‐to‐volume registration was used for reconstruction of 3D isotropic (0.85 mm) volumes of the fetal body followed by semi‐automated lung segmentation. 3D TLV were compared to traditional 2D‐based volumetry. Abnormal cases referenced to a normogram produced from 100 normal fetuses whose TLV was calculated by DSVR only. Results Deformable slice‐to‐volume registration‐derived TLV values have high correlation with the 2D‐based measurements but with a consistently lower volume; bias −1.44 cm3 [95% limits: −2.6 to −0.3] with improved resolution to exclude hilar structures even in cases of motion corruption or very low lung volumes. Conclusions Deformable slice‐to‐volume registration for fetal lung MRI aids analysis of motion corrupted scans and does not suffer from the interpolation error inherent to 2D‐segmentation. It increases information content of acquired data in terms of visualising organs in 3D space and quantification of volumes, which may improve counselling and surgical planning.

� Congenital diaphragmatic hernia (CDH) and congenital lung lesions (CLL) are prognosticated with ultrasound-based measurements of the fetal lung in a single dimension; however true volumes may provide greater sensitivity for high risk cases. Current use of magnetic resonance imaging (MRI) to calculate fetal lung volumes is limited as two-dimensional segmentation is labour intensive and risks interpolation and motion-corruption errors.

What does this study add?
� Three-dimensional lung volumes can be computed from deformable slice-to-volume registration (DSVR) 3D reconstructions and highly correlate with traditional 2D-derived volumes. DSVR-derived volumes, however, should be more reliable owing to higher resolution and semi-automated calculations that do not rely on interpolation between slices on motion-corrupted stacks.

| INTRODUCTION
The antenatal work-up of fetal thoracic anomalies increasingly includes fetal magnetic resonance imaging (MRI) in recent years.
Fetal MRI aids in cases of diagnostic uncertainty by detailed delineation of soft tissues within the fetal chest, differentiating lung, bowel and liver more easily than ultrasound, 1 5 Discrepancies between 2D ultrasound and MRIderived TLV measurements have been shown to be principally related to the failure to account for the contribution of the ipsilateral lung. 6 Furthermore, ultrasound-based imaging struggles with maternal habitus and fetal positioning, as well as extremes of liquor volume. 3D ultrasound lung volume measurements have been shown to be more difficult than by MRI, mainly because the most hypoplastic lung cannot be properly visualised. 7 While US-based measurements are undoubtedly the bedrock upon which assessment of fetal lung anomalies should be based, high volume fetal medicine centres have suggested that MRI-calculated lung volumes may be a more accurate predictor of survival. 8 This has not been formally proven and there is no standardised methodology for the use of MRI in these cases which limits the compilation of large datasets necessary for further study. 9 Modern clinical fetal MRI protocols (primarily based on single shot turbo spin echo [ssTSE] sequence) allow fast acquisition of individual 2D slices that "freeze" fetal position in time. These slices have sufficiently high image and contrast resolution for diagnostic purposes in cases where fetal motion may previously have limited the information available. However, the misalignment between individual slices leads to corruption of volumetric information and loss of structural continuity within a slice stack. Therefore, output MRI stacks are termed "motion corrupted". In general, to achieve a suf- The recently proposed deformable slice-to-volume registration (DSVR) method 10 is used for reconstruction of high-resolution (e.g., 0.8 � 0.8 � 0.8 mm) isotropic 3D images of fetal body from multiple low-resolution (e.g., 1.25 � 1.25 � 1.25 mm) motion-corrupted stacks. In DSVR method, one of the low-resolution stacks is selected as an initial target space and it is then registered to each of the slices using nonlinear free form deformation registration. This is followed by super-resolution reconstruction (SR) of the 3D image from the registered slices. The full pipeline includes three interleaved SR and SVR steps. The resulting reconstructed images provide detailed 3D volumetric information and can be reoriented in any plane. This facilitates accurate 3D segmentations of fetal lungs and other organs allowing true volumetric analysis.
In this work, we sought to explore how 3D DSVR-derived lung volumes would compare to those calculated by conventional manual 2D segmentation in cases of fetal thoracic anomalies, with a comparison to normal control cases.

| Case selection
The investigated cohort for 2D versus 3D TLV assessment included eight cases with abnormal lungs and eight control cases from the iFIND project [https://www.ifindproject.com] and clinical fetal cardiac MRI databases. The abnormal cases included six cases of CDH (one with concomitant BPS), and a further two cases of CLL (both macrocystic CPAM). The eight control cases were healthy control pregnancies and spanned approximately the same gestational range as the abnormal cases (22-32 weeks gestational age). In addition, 100 normal cases imaged as healthy control participants for research purposes were used for the general assessment of 3D DSVR-derived fetal lung volumetry growth chart. These cases were selected by stratified random sampling in order to produce the widest possible range of gestational ages. Cases with extreme motion (i.e., >45°r otation in the body stacks) were excluded as their reconstruction is known to be challenging and may be prone to error, in practice this amounted to less than 5% of all cases in our database. All datasets used in this research were collected and processed subject to the informed consent of the participants.

| Data
Each fetus was scanned using T2-weighted ssTSE sequence, producing between 6 and 10 stacks with minimum of 5 orientations (orthogonal planes with respect to the uterus, fetal trunk and brain regions), with varying degrees of motion corruption (none severe as outlined above).
No maternal breath hold or sedation was used during the acquisition. In all cases, images were acquired on an Ingenia 1.5 T (Philips) system using    12 The resultant curve was depicted graphically along with previously published MRI-generated fetal lung volume normograms. 12,13 As this was a pilot study, there was no power calculation performed and no further statistical analysis was deemed appropriate to perform.

| RESULTS
The examples of 2D versus 3D lung segmentations for 4 cases are depicted in Figure 2; motion corruption of the axial-plane stack can clearly be appreciated in the coronal plane, with the coronal section of the DSVR shown alongside each case. In should be noted that the employed thin slice acquisition protocol with 1.25 slice spacing provides denser sampling of the lungs thus ensuring higher spatial coverage. However, it is also more susceptible to motion artifacts in comparison to the conventionally used 3 mm spacing due to the longer scanning times. In comparison, DSVR was able to produce clear 3D images even from severely motion-corrupted stacks ( Figure 2C) and the resultant images can be examined in any plane and at a higher resolution than that of the original input stack (Figure 1, Video 1). The resultant segmentation of fetal lungs demonstrates a smooth outline, with normal cases depicting sufficient detail even to appreciate the lingula ( Figure 2A) or a systemic feeding vessel in cases of CLL ( Figure 2B). Comparison of the volumes generated from 2D and 3D segmentations is given in Figure 3; results correlated strongly, however a consistently lower measurement was made with 3D segmentation (Bland-Altman: Bias −1.44 cm 3 [95% confidence limits −2.63 to −0.24]). This was felt to be due to the enhanced ability to exclude structures of the pulmonary hilum in 3D reconstructions, as well as an inherent reduction in interpolation error by accommodating motion artefact.

F I G U R E 2
Having defined the fidelity of DSVR-derived 3D TLV, we proceeded to produce a normal curve based upon 100 healthy cases undergoing fetal MRI as normal controls for research, presented in The manually segmented 2D lung volumes and corresponding 3D DSVR-derived lung volumes are depicted graphically in Figure 5 with reference to the DSVR-derived normal curve. Table 1

| DISCUSSION
This pilot study demonstrates that MRI of the fetal thorax can be processed by 3D DSVR reconstruction, and subsequent lung volumetry can be performed with volumes generated that highly F I G U R E 3 2D manual segmentation derived total lung volume (TLV) plotted against deformable slice-to-volume registration (DSVR)-derived 3D TLV. Bland-Altman bias −1.44 cm 3 . For reference the line y = x is also drawn (i.e., perfect match of 2D and 3D volumes) F I G U R E 4 Normal total lung volumes as calculated by DSVR; 100 normal cases were used to generate a curve with the equation TLV (ga) = 0.00028ga 3.57 , with an R 2 = 0.72. For reference the curves suggested by Cannie 13 and Meyers 12 are included on the same graph correlate to the "gold standard" 2D-manual segmentation derived volumes currently used in clinical practice for O/E TLV assessment.
Furthermore, 3D lung segmentations potentially provide more accurate TLV estimation since they minimise the errors on the conventional 2D-based approach: segmentation of 2D stacks and subsequent volumetric measurements inherently cannot be regarded as "true" to the subject, since a degree of motion corruption is always present in these acquired images and the slice-to-slice measurements will require interpolation in order to produce workable volumes. The Bland-Altman analysis of the quantitative results as well as the general visual inspection of 2D versus 3D segmentations suggest that using 3D DSVR-based assessment may minimise these errors as the resultant images are not only of a higher resolution, but also aprovide continuous 3D information does not affected by interpolation. This means that small structures of the thorax, such as hilar vessels, can be reliably excluded in all cases, producing volume measurements far closer to that of the patient. This potentially resolves one of the current limitations to the validity of MRI derived volumes in predicting outcomes as reliably as LHR. 14 Three dimensional images produced from MRI of the fetus have been conclusively demonstrated to enhance diagnostic capabilities of MRI in assessment of the fetal brain. 15 This pilot series demonstrates that DSVR method provides the means for accurate calculation of lung volumes for both normal and abnormal cases. This is especially F I G U R E 5 Total lung volumes of normal and abnormal cases computed from 3D deformable slice-to-volume registration (DSVR)-derived and 2D manual slice-wise segmentations with reference to our DSVR-derived normogram

Case
Uss assessment -633 useful for CDH cases for more accurate evaluation of the observed to expected lung volume ratios and the volume of hypoplastic lung. The recent advances in image processing methods reduce variability in segmentation, 16 further improving the reliability of the volumetry results. We anticipate that MR based lung volumetry may soon be validated as a reliable and more accurate prognostic indicator than the current use of the LHR; since it would invariably more closely predict true TLV. Furthermore, the ability to outline lesion volumes in cases of CLL will likely produce a prognostic marker more faithful to true lesion size than the currently utilised CVR where cross sectional areas are utilised to estimate lesion volume relative to fetal size. The consistency of such a method will certainly deliver on the recent plea for consistency in prognostication of these cases. 9 It should be noted that although we have provided lesion volume proportionate to TLV, it would be possible to segment any aspect of the fetus and derive a corresponding volume ratio proportional to this as well as the relative position of the organs (e.g., the degree of liver herniation or position of the stomach in CDH cases). Furthermore, automation of these 3D measurements using deep learning tools will potentially allow unbiased analysis for large cohorts.

3D-Observed TLV (CM 3 ) 2D O/E TLV 3D O/E TLV
Recently, it has also been demonstrated that DSVR reconstructed MRI images may provide additional diagnostic information for complex fetal anomalies. 1 The illustrative cases included within this dataset also show that DSVR will produce datasets that are more accessible to the clinician by resolving motion corruption with a resolution superior to that of the original input stacks. This will certainly aid in prenatal counselling and planning of the operative approach in neonatal thoracic surgery; the reliable depiction of systemic feeding vessels in BPS, combined with US Doppler assessment may reduce the reliance on early postnatal CT imaging. 17 Regarding longitudinal assessment of CDH or CLL, it should be recognised that the repeated use of MRI is considerably more expensive, time consuming and lengthy compared to an expertperformed ultrasound. Therefore, we anticipate that there will continue to be a role for ultrasound in the monitoring of such conditions to delivery. Indeed, serial CVR assessment has recently been shown to be effective in predicting a need for perinatal intervention in cases of CLL. 18

| CONCLUSIONS
This pilot study demonstrates the potential of a 3D reconstruction platform such as DSVR in the analysis of the fetal lung, with an emphasis on accuracy and reliability in image derived volumetry. We anticipate that MR based lung volumetry may soon be validated as a reliable and more accurate prognostic indicator than the current use of the LHR.
Our current work focuses on deep learning methods in order to provide reliable volumetric assessment for a wide range of anomalies as well as optimisation of different MRI acquisition protocols. This will allow integration of both DSVR and volumetry pipelines in different clinical centres that employ fetal MRI on a regular basis.